Editor integration: LSP tools and ACP server#

Settled 2026-08-10 with the owner. Of the editor-integration gap rows in the feature matrix, two are worth building, one is a matrix re-audit, five are deliberate non-goals.

LSP tools (first)#

The beneficiary is the agent, not an editor. OpenProgram runs language servers as background analysis processes and exposes their answers as agent tools:

• lsp_diagnostics(file) — compiler-grade errors and warnings for a file, available the moment the agent finishes an edit, without running tests.
• lsp_references(file, line, column) — every real call site of a symbol; grep misses dynamic calls and false-positives on same-named strings.
• lsp_definition(file, line, column) — the symbol's true location across files and packages.

Scope: Python (pyright) and TypeScript (typescript-language-server) — the two languages this repo actually contains. One server per language per workspace, started on first use, cached, shut down with the session. A missing server binary degrades to a clear "unavailable: install X" tool result, never a crash; the tools stay registered so the model learns what is possible.

Relation to CodeGraph: CodeGraph is the pre-built symbol index for understanding a codebase; LSP adds what an index cannot — live diagnostics against the current unsaved state of the working tree.

ACP server (second)#

ACP (Agent Client Protocol) is the editor-agnostic standard for editors driving an external agent. Implementing the server side maps three matrix rows at once: editors like Zed drive OpenProgram sessions directly (被IDE经ACP驱动), the editor ships selection and open-file context with each request (编辑器选区与打开文件进上下文), and the editor becomes an entry point without any self-built extension (官方IDE扩展, partially). One stdio protocol adapter over the existing session/tool loop. Starts after LSP lands.

Matrix re-audit#

桌面GUI控制 is marked absent, but GUI-Agent-Harness is exactly that capability, installed as a program. Re-audit the row against the matrix's own evidence rules (installed-harness capability vs built-in).

Non-goals#

Self-built IDE extension (ACP replaces it), inline completion (a different product), inline chat / CodeLens / status-bar actions (depend on a self-built extension), terminal shortcut auto-configuration (return too small). These rows stay open in the matrix by choice.

Implementation status#

LSP tools have landed. The client lives in openprogram/lsp/ (JSON-RPC over stdio, one server per language per workspace, started on first use and shut down at exit); the three tools live in openprogram/functions/tools/lsp/ and are documented for users in Language server tools. Server processes go through the same sandbox-wrapping entry as every other child, so sandbox.mode applies to them unchanged.

The ACP server has landed. openprogram/acp/ holds the JSON-RPC ndjson transport (jsonrpc.py) and the protocol mapping (server.py), reached via openprogram acp. It implements protocol version 1: initialize, session/new, session/load, session/prompt and session/cancel from the editor, session/update and session/request_permission back to it. Users configure it in Editor (ACP).

The adapter sits on dispatcher.process_user_turn, the same non-HTTP entry the webui thread, sub-agents and the task runner use, so tool gating, authority and persistence are shared rather than reimplemented. Turns carry local_owner_authority() — an editor on the same machine is the owner speaking interactively, which is what makes approval.request available at all. Approvals reach the editor by subscribing to the question.asked event and answering through resolve_question_and_broadcast, so the existing gate, its rules, and the "always allow" rule-persistence path are untouched.

Deliberately unimplemented: session/set_mode and session/set_model (the agent id and permission mode are process flags), the editor's fs/* and terminal/* client capabilities (OpenProgram's own tools already read and write the working tree), and MCP servers passed in session/new (OpenProgram uses its own MCP configuration).

Still open: the matrix re-audit of the GUI row.